Colour is requested as a mood and delivered as an approximation

A colour name is a region, not a value

Ask for a deep teal background and you will get something teal, at a saturation and lightness that varies from one generation to the next. Ask for a specific brand colour by its code and the result is often close, occasionally not close at all, and never reliably exact across a set of images.

This surprises people because colour feels like the most objective thing in a picture. It is objective as a value and vague as a word, and a prompt is words. Every colour term covers a range that different sources use differently, and the output lands somewhere inside that range.

The practical consequence is that colour should never be the thing you are relying on the prompt to nail. If a particular shade is a requirement rather than a preference, that requirement is going to be satisfied somewhere other than the generation step.

Relationships are more controllable than values

What does work is describing colour relationally. A single warm accent against a desaturated background. Two dominant colours and nothing else. Everything within a narrow range of blues except one object. Low saturation throughout. These are structural descriptions and they survive from one image to the next far better than a named shade.

The reason is that a relationship constrains the whole picture at once, whereas a colour name constrains one region and leaves everything around it free. A prompt that specifies the accent colour and says nothing about the rest will produce an inconsistent set even when the accent is correct every time.

Lighting descriptions carry colour as well, and often more predictably than colour words do. Late afternoon light, overcast, fluorescent interior, single warm lamp in a dark room — each of these implies a palette, and it implies it consistently because it is describing a physical situation rather than picking a swatch.

Fix exact colours after generation, not during

For anything that has to match — a brand palette, a series that will sit on one page, an illustration that has to work against a known background — the workable order is to generate with the composition and mood you want, then correct the colour in an ordinary image editor.

This isn’t a workaround, it is the normal division of labour. Colour correction is an old, precise, well-understood operation. Selecting a region and replacing its hue, adjusting a whole image towards a target temperature, or mapping a picture onto a fixed palette are all things that ordinary tools do exactly, repeatably, and in seconds.

It also means you can stop trying to persuade the generator. Three additional attempts at the right shade cost more time than one adjustment, and the adjustment produces the actual value rather than a near miss that will drift again in the next image.

The exception is when the colour is doing something physical in the scene — coloured light falling on an object, a reflection, a translucent material. Those cannot be corrected regionally afterwards without the result looking painted, and they are worth getting from the generation.

Watch what colour does to everything else

Colour words in a prompt rarely act alone. Asking for something in muted earth tones tends to bring a whole associated look with it — soft light, film texture, a particular kind of composition — because the phrase is more strongly associated with a genre of photograph than with a set of hues.

This is the same trap as style words generally, and it shows up most in the difference between what you asked for and what changed. Add a colour instruction to a working prompt and check the composition afterwards, because the two aren’t as separable as they should be.

If the palette shift drags the picture somewhere you didn’t want, separate the request: describe the light physically, describe the palette relationally, and remove the evocative phrase that was carrying both plus a mood you never asked for.

Screens, print and the limits of matching

A last practical caution, and it has nothing to do with generation. Colours that match on your screen may not match in print, on another display, or in a document viewed under a different colour setting, and this defeats a surprising amount of careful work.

For anything that will be printed or that has to sit against fixed brand assets, the colour decision belongs to whoever handles the final file, in the colour space that file uses. Generated images enter that process as material, exactly as photographs do.

Which is the sensible way to think about the whole question. Get composition, subject and light from the generation, because those are hard to obtain any other way. Get exact values from the tools that were built to produce exact values.